model-release-review

name	model-release-review
description	Audits a commercial-photography catalog for entries that require a signed model release. Four-tier classification (HARD-FLAG / SOFT-FLAG / OK / EXEMPT) keyed to identifiability-to-the-general-public, not facial recognition alone. Stricter bar for children (parental release) and workers at workplaces. Mandatory 2048px-long-edge downscale before viewing any image — pixels falsify description-only assumptions every time. Pairs with property-release-review for the rights side of the same catalog. USE WHEN model release review, person release, can I sell this print, identifiable person audit, candid model audit, child in catalog, worker in catalog, portrait release audit. NOT FOR property/architecture/artwork rights (use property-release-review).
effort	medium

🚨 MANDATORY FIRST ACTION: Downscale Every Image to 2048px Long Edge

Before reading or classifying ANY image, downscale to 2048px on the long edge. Process only the downscaled copy. If resize fails, HALT and prompt the user.

import sharp from "sharp";
await sharp(srcPath)
  .resize({ width: 2048, height: 2048, fit: "inside", withoutEnlargement: true })
  .jpeg({ quality: 82, mozjpeg: true })
  .toFile(outPath);

magick "$FILE" -resize "2048x2048>" "$OUT" is an acceptable ImageMagick fallback.

Output location: a scratch dir scoped to the active audit (e.g. <workdir>/downscaled/). Originals are never modified. If any resize fails, stop and ask — do NOT skip and continue.

When identifiability remains borderline even at 2048px: 2048px is the maximum resolution for batch auditing and resolves nearly all cases. When a face is small enough, far enough, occluded enough, or turned enough that you cannot confidently assign HARD-FLAG vs OK even at full downscale resolution, do not guess — escalate to human judgment. Wrongly clearing a release flag (letting a person be sold without consent) is the expensive direction to be wrong in. Document the reason ("face too small", "partially obscured", "half-turned") and mark for escalation.

This rule exists because description-only judgments (carry-over notes from prior sessions, captions, IPTC metadata) lie. Pixels don't. The gate has demonstrably caught description-only HARD-FLAGs that the actual frame falsified — back-turned subjects mislabeled as frontal, glass-distorted faces mislabeled as identifiable.

Intake Sequence (Step 4 of 5)

For new-photo intake, this skill is step 4 of 5 in the pipeline:

1. photo-metadata-helper — metadata, naming, subject-name embed
2. quality-review — technical, editorial, and print-readiness gate
3. property-release-review — depicted-object audit
4. model-release-review (this skill) — depicted-person audit
5. reviewed-photo-publish — catalog entry, sales-platform listing, remove from intake queue

If quality-review failed a photo or property-release-review flagged it as Bucket 1 (remove entirely), skip model review for that photo — the upstream concern is already dispositive. Publish is a distinct downstream step, never collapsed into this skill.

For audit-only passes on an already-published catalog, this skill can be invoked standalone.

The Question

For each non-exempt catalog entry: does the image require a signed model release before commercial sale?

Answer = HARD-FLAG (yes) / SOFT-FLAG (judgment call) / OK (no) / EXEMPT (skipped per catalog owner's declaration).

Identifiability Standard

The legal bar is identifiable to the general public, NOT identifiable to close friends. Personality / privacy rights protect against being recognized by the public from the image alone.

This is the most common over-flagging mistake: assuming a four-attribute "fingerprint" (coat + bag + hair + hat) of generic items in a public place makes someone identifiable. It almost never does — the bar is what a stranger could match against social media or someone's general appearance.

Factor	Pushes toward HARD-FLAG	Pushes toward OK
Face	Visible, in focus, frontal or 3/4	Back-turned, silhouette, deep shadow, glass distortion
Distance	Subject occupies significant frame	Small/distant figure, not the subject
Subject vs incidental	Person is what the photo is about	Person is compositional element only
Clothing	Branded, uniformed, singular	Generic, mass-retail
Body / posture	Distinctive (tattoos, gait, public-known build)	Generic
Setting	Workplace (worker depicted at work)	Public space, anonymous context

Strict cases — always HARD-FLAG even with face partially obscured:

• Children — parental release required regardless of identifiability nuance
• Workers at workplace — cook at food stall, driver in cab (when face IS visible — glass distortion is dispositive otherwise), performer on stage. Selling prints of someone doing their job is a stricter standard than editorial use.

Permissive cases — usually OK even though a person is in frame:

• Pure back-turn + generic clothing + public space (archetypal/anonymous framing)
• Heavy glass distortion or reflection over the face
• Silhouettes against strong backlight
• Crowd shots where no individual is the subject
• Distance great enough that features cannot be discerned at print scale

Tiers

Tier	Action	Trigger
🟥 HARD-FLAG	Release required before sale	Identifiable face/features, person is subject; OR child; OR worker at workplace
🟧 SOFT-FLAG	Judgment call — surface to user	Framing/distance ambiguous; reasonable people could disagree
🟩 OK	No release needed	Unidentifiable per the table above
🚫 EXEMPT	Skip entirely	Categories the catalog owner has declared exempt (e.g. macro, abstract, in-studio)

Portraits — Audit, Not New Ask

Posed/commissioned portraits are HARD-FLAG by series convention, but a working photographer typically already has releases on file. Treat the portraits pass as audit-confirmation (list them so the owner can verify the signed release exists for each subject), not as a flag for new acquisition.

Calibrated Debate, Not Capitulation

When the user challenges a flag, debate it honestly:

1. Re-apply the 2048-gate (downscale + view the actual pixels)
2. State the strongest counter-argument you can find
3. Test it against the identifiability standard
4. Concede when the argument is dispositive; name residual risk when not

The downscale gate exists because pixels routinely falsify carry-over notes — view first, opine second.

Workflow

1. Scope — read the catalog, count entries, separate exempt category(ies) from non-exempt
2. Preliminary classification from descriptions/series — portraits HARD-FLAG (audit), architecture-only OK by default, anything with humans → needs visual
3. Visual verification (2048-gate MANDATORY) — downscale every non-OK candidate, view, classify
4. Spot-check — pick ≥2 OK-by-description entries, downscale + view, confirm
5. Structured report — HARD-FLAG / SOFT-FLAG / OK / EXEMPT tiers; every HARD/SOFT has one-line reason; OK count tallied by exclusion
6. Surface decision points — dispositions per HARD-FLAG (release file lookup, delist, or remove); flag any portraits where release status is uncertain

Dispositions (mirror of property-release-review)

Bucket	Action	When
Bucket 1	Remove entirely	Release impossible AND image too risky to keep even unsold
Bucket 2	Delist from sales platform; keep in portfolio	Release impossible but image acceptable as portfolio
Bucket 3	Keep on sale, release confirmed on file	Release exists, audit-confirmed

Implementation depends on your catalog architecture — see property-release-review for the generic remediation pipeline.

Originals of removed photos go to an archive dir, not deletion. Restoration must remain possible.

Distinction from property-release-review

Concern	This skill	Sibling
Faces / persons	✓	—
Buildings / interiors	—	✓
Permanent artwork (sculpture, murals)	—	✓
Branded venues, trademarks	—	✓
Child protection	✓	—
Worker-at-workplace	✓ (model release)	partial (where branded uniform / workplace property)
Freedom-of-Panorama analysis	—	✓
Owner-declared exempt category	✓	✓
2048-gate	✓	✓
Debate-don't-capitulate	✓	✓

When a photo trips both — escalate via the sibling skill first (property) since trademark/permanent-artwork concerns are usually dispositive of the sale decision regardless of model release.

Lessons Encoded From Live Reviews

• Pixels falsify descriptions. Description-only HARD-FLAGs routinely break under inspection — back-turned subjects misremembered as frontal, glass-distorted faces misremembered as visible. The downscale gate exists because of these failure modes.
• Glass distortion is dispositive. Heavy slanted/reflective glass over a face = OK, not HARD-FLAG, even if the person is a worker at workplace.
• Back-turn + generic clothing + public space = OK. Even with a four-attribute "fingerprint" (coat, bag, hat, hair), the public-identifiability bar is rarely met.
• A model-release OK doesn't restore a photo already off-sale for property/trademark reasons. The two reviews are independent; resolve property first.

Gotchas

• Do not enumerate the exempt category in the report — it's EXEMPT en bloc.
• Do not flag based on description alone — always downscale + view first.
• Do not auto-bucket HARD-FLAGs — present to the catalog owner and let them decide.
• Children require parental release; "child looks happy / parents nearby" is not a substitute.
• Worker-at-workplace HARD-FLAG persists even if the worker is part of a brand the photo glorifies — selling prints of someone doing their job is not editorial.

When to invoke

• "review the catalog for model releases"
• "do I need a release for [photo]"
• "model release audit"
• "can I sell this print" (person in frame)
• "person release", "candid release"
• "review portraits for release on file"
• Adding catalog entries that contain people other than commissioned subjects